Intelligent switch cabinet intelligent control method and control system

Through the intelligent control method of intelligent switch cabinet, the neural network uses data flow identification to process the monitoring and control data flow of switch cabinets, which solves the problem of inefficiency of traditional control methods, realizes accurate and reliable control of switch cabinets, and improves the power system stability.

CN118584885BActive Publication Date: 2024-12-03GUANGDONG MILLER ELECTRIC CO LTD
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202411065224.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-08-05
Publication Date
2024-12-03
Estimated Expiration
2044-08-05

AI Technical Summary

Technical Problem

The traditional switch cabinet control method is inefficient, prone to misoperation, lacks real-time data analysis and prediction capabilities, affecting the reliability and safety of the power system. The existing intelligent transformation has failed to form a systematic intelligent control solution.

Method used

Using intelligent switch cabinet intelligent control method, the target monitoring data flow and control data flow are obtained, the relevant data flow is determined, the fused data flow is formed, and the data flow is loaded into the data flow identification neural network, and the vector is characterized by using the preset data flow characterization vector. Fusion, calculate the control trigger probability, and achieve accurate and reliable control of the switch cabinet.

Benefits of technology

It improves data processing speed, accurately extracts the encoding characteristics of data flow, enhances the control decision support of the switch cabinet, and ensures the stable operation of the power system.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118584885B_ABST
    Figure CN118584885B_ABST
Patent Text Reader

Abstract

The present application provides an intelligent control method and control system for an intelligent switch cabinet, which performs data stream screening on the control data stream set of the target switch cabinet to obtain the relevant data stream set of the target monitoring data stream, which can prevent direct encoding of large-scale data streams, reduce computing overhead, and increase data processing speed. The neural network is obtained based on the debugging of the control data stream set of the prior switch cabinet, and can accurately extract the coding features of the data stream set. The fused data stream set obtained based on the relevant data stream set and the target monitoring data stream is loaded into the target data stream identification neural network for encoding, and accurate features can be obtained. The neural network is based on the initial data stream characterization vector searched for the target monitoring data stream and the relevant data stream set, and the target monitoring data stream and the relevant data stream set are collaboratively encoded. Based on the data stream set characterization vector obtained by encoding, the characteristics of the target switch cabinet on the target monitoring data stream can be accurately characterized.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to an intelligent control method and control system for an intelligent switch cabinet. Background Art

[0002] In the power system, the switchgear is a key power distribution and control equipment, and its operation stability and safety are crucial to the stable operation of the entire power system. However, with the continuous development and complexity of the power system, the traditional switchgear control method has gradually exposed many shortcomings. Traditional switchgear mainly relies on simple protection, which is not only inefficient, but also prone to misoperation, increasing the risk of operation. In addition, traditional switchgear lacks the ability to deeply analyze and predict real-time operation data, making it difficult to detect and respond to potential faults in a timely manner, affecting the overall reliability and safety of the power system. In the prior art, although there have been some attempts to intelligently transform switchgear, these attempts are often limited to the automation upgrade of local functions and fail to form a systematic intelligent control solution. For example, some systems monitor the internal parameters of the switchgear by adding sensors, but lack efficient data processing and analysis capabilities, and cannot extract valuable information from massive data for control decisions. How to achieve real-time and accurate monitoring of the internal parameters of the switchgear has become an urgent problem to be solved. Summary of the invention

[0003] In view of this, the present application provides an intelligent switch cabinet intelligent control method and control system.

[0004] The technical solution of the embodiment of the present application is implemented as follows: On the one hand, the present application provides an intelligent control method for an intelligent switch cabinet, the method comprising: obtaining a target monitoring data stream and a control data stream set of a target switch cabinet; determining a related data stream set of the target monitoring data stream in the control data stream set of the target switch cabinet; obtaining a fused data stream set based on the related data stream set and the target monitoring data stream, and loading the fused data stream set into a target data stream identification neural network; the target data stream identification neural network includes a preset data stream characterization vector set, and the target data stream identification neural network is obtained by debugging the control data stream set of the prior switch cabinet; based on the target data stream identification A neural network searches for a preset data stream characterization vector set corresponding to each data stream in the fused data stream set in the preset data stream characterization vector set, performs characterization vector fusion on each searched preset data stream characterization vector, and obtains a data stream set characterization vector of the fused data stream set; the data stream set characterization vector is used to determine the control trigger probability of the target switch cabinet on the target monitoring data stream; on the other hand, the present application provides a control system, including a CPU main control chip (domestic) and a FLASH memory chip (domestic), etc. The FLASH memory chip can run a computer program on a processor, and the steps in the above method are implemented when the CPU main control chip executes the program.

[0005] The beneficial effects of the present application include at least: the intelligent control method and control system of the intelligent switch cabinet provided by the embodiment of the present application obtains the target monitoring data stream and the control data stream set of the target switch cabinet; determines the relevant data stream set of the target monitoring data stream from the control data stream set of the target switch cabinet; obtains the fused data stream set based on the relevant data stream set and the target monitoring data stream, and loads the fused data stream set into the target data stream identification neural network; the target data stream identification neural network includes a preset data stream characterization vector set, and the target data stream identification neural network is obtained by debugging the control data stream set of the prior switch cabinet; based on the target data stream identification neural network, searches for the preset data stream characterization vector set corresponding to each data stream in the fused data stream set in the preset data stream characterization vector set, fuses the characterization vectors of each preset data stream characterization vector searched, and obtains the data stream set characterization vector of the fused data stream set; the data stream set characterization vector is used to determine the control trigger probability of the target switch cabinet on the target monitoring data stream. Based on this, the control data stream set of the target switch cabinet is screened for data streams to obtain the relevant data stream set of the target monitoring data stream, which can prevent direct encoding of large-scale data streams, effectively reduce computing overhead, and increase data processing speed. The target data stream identification neural network is obtained by debugging the control data stream set of the prior switch cabinet, and can accurately extract the coding features of the data stream set. Then, the fused data stream set obtained based on the relevant data stream set and the target monitoring data stream is loaded into the target data stream identification neural network for encoding, and accurate features can be obtained. The target data stream identification neural network is based on the initial data stream characterization vector searched for the target monitoring data stream and the relevant data stream set, and performs collaborative encoding on the target monitoring data stream and the relevant data stream set. Based on the encoded data stream set characterization vector, the characteristics of the target switch cabinet on the target monitoring data stream can be accurately characterized. Afterwards, the control trigger probability of the target switch cabinet on the target monitoring data stream can be determined based on the data stream set characterization vector, making the control of the switch cabinet more accurate and reliable. BRIEF DESCRIPTION OF THE DRAWINGS

[0006] Figure 1 A schematic diagram of the implementation process of an intelligent switch cabinet intelligent control method provided in an embodiment of the present application;

[0007] Figure 2 A hardware entity schematic diagram of a control system provided in an embodiment of the present application. DETAILED DESCRIPTION

[0008] The embodiment of the present application provides an intelligent control method for an intelligent switch cabinet, which can be executed by a CPU main control chip of a control system, wherein the control system refers to equipment with data processing capabilities such as an intelligent distributed DTU and a feeder automation FTU.

[0009] Figure 1 A schematic diagram of the implementation process of an intelligent switch cabinet intelligent control method provided in an embodiment of the present application is shown in FIG. Figure 1 As shown, the method includes: Step S10: acquiring a target monitoring data flow and a control data flow set of a target switch cabinet.

[0010] In the embodiment of the present application, the target monitoring data stream refers to those data streams that have a key impact on the operating state or environmental conditions of the switch cabinet, such as the voltage and current data of the power system, or environmental parameters such as ambient temperature and humidity. These data are collected in real time through the sensor network and recorded in a specific format (such as time series data). For example, in the power system, the target monitoring data stream may include voltage and current values ​​collected once a minute, and these values ​​constitute a time series that reflects the real-time operating state of the power system. Next, the control system obtains a set of control data streams of the target switch cabinet, which contains all data stream records that have triggered the control behavior of the switch cabinet in history. These control data streams are consistent with the target monitoring data streams in format and collection content, but the key is that they directly lead to certain control actions of the switch cabinet, such as the closing or opening of the circuit breaker, protection actions, etc. These control data streams also come from the sensor network or the monitoring system of the switch cabinet itself, which records the time, conditions and execution results of the control action. For example, when an abnormal increase in voltage is detected, the control system can record a control data stream indicating that the opening operation of the circuit breaker is triggered at a certain point in time to prevent equipment damage.

[0011] It should be noted that although the control data stream and the target monitoring data stream are similar in form, their meanings and uses are completely different. The target monitoring data stream is used for real-time monitoring and predictive analysis, while the control data stream reflects the actual control logic and response mechanism of the switchgear. When acquiring these data streams, the control system ensures their integrity and consistency so that subsequent steps can accurately perform data processing and analysis.

[0012] Step S20: determining a data flow set related to the target monitoring data flow in the control data flow set of the target switch cabinet.

[0013] Specifically, the control system deeply analyzes the characteristics of the target monitoring data stream, including its data pattern, change trend, and key feature points. Taking voltage monitoring in a power system as an example, the target monitoring data stream may be a series of continuously recorded voltage values ​​that fluctuate over time, reflecting the real-time status of the power system.

[0014] Next, the control system compares and analyzes the target monitoring data stream with the control data stream set of the target switchgear. The control data stream set contains all the data stream records that have triggered the switchgear control behavior in history. These records also contain time series data of key parameters such as voltage and current. The goal of the control system is to find control data streams that are similar in characteristics or related in behavior to the target monitoring data stream.

[0015] To achieve this goal, the control system can use a variety of data analysis techniques, including but not limited to similarity calculation, cluster analysis, time series analysis, etc. For example, the control system can calculate the spatial similarity coefficient (such as Euclidean distance, cosine similarity, etc.) between the target monitoring data stream and each control data stream to quantify the degree of similarity between them. In the power system scenario, if a control data stream also records a similar fluctuation pattern when the voltage rises abnormally, then this control data stream is likely to be associated with the target monitoring data stream. In addition, the control system can also use machine learning models to assist in this task. For example, it can train a classification model based on a neural network or a decision tree, which can learn the feature representations of different data streams in the control data stream set and identify data stream patterns that are similar or associated with the target monitoring data stream. During the training process, the control system uses labeled historical data to guide the model's learning process to ensure that the model can accurately capture key features and behavior patterns.

[0016] Finally, through this step, the control system can determine a set of data streams associated with the target monitoring data stream. Each data stream in this set may reflect the future change trend or potential risk point of the target monitoring data stream to some extent. Based on this set, the control system can further perform predictive analysis, risk assessment or control strategy optimization, thereby providing more accurate and reliable decision support for the intelligent control of the intelligent switch cabinet.

[0017] For example, in a power system, if the target monitoring data stream shows that the voltage is gradually increasing, the control system can find all the data stream records that have triggered control actions when the voltage increased in history through step S20. These records may include circuit breaker opening operations, protection action status, etc., which together constitute a data stream set associated with the target monitoring data stream. By deeply analyzing the data stream characteristics and behavior patterns in this set, the control system can predict the possible future voltage change trend and formulate corresponding control strategies in advance to ensure the stable operation of the power system.

[0018] Step S30: Based on the relevant data flow set and the target monitoring data flow, a fused data flow set is obtained, and the fused data flow set is loaded into the target data flow identification neural network; the target data flow identification neural network includes a preset data flow characterization vector set, and the target data flow identification neural network is obtained based on the debugging of the control data flow set of the prior switch cabinet.

[0019] Specifically, the control system fuses the target monitoring data stream itself with the result of step S20, that is, the set of related data streams of the target monitoring data stream, to form a new fused data stream set. This set not only contains the real-time information of the target monitoring data stream, but also integrates the control data stream characteristics associated with it in history, thereby providing a more comprehensive data perspective. Taking voltage monitoring in the power system as an example, the target monitoring data stream may be the voltage value and its changing trend at the current moment, while the set of related data streams may contain data records of control behaviors (such as line grounding faults, short circuit faults, etc.) triggered in similar voltage changes in the past. The control system aligns these data records with the target monitoring data stream in time series to form a fused data stream set containing rich historical information and real-time data.

[0020] Next, the control system loads this fused data stream set into the target data stream identification neural network. This neural network is specially designed to identify and analyze complex patterns and features in the switchgear control data stream. It contains a set of preset data stream representation vectors, which are obtained by analyzing the control data stream sets of a large number of a priori switchgears (i.e., the switchgears used in the training phase). These vectors can efficiently characterize the key features of different data streams, such as the voltage fluctuation range, change rate, duration, etc., thereby providing strong support for subsequent prediction and control. In the implementation process, the target data stream identification neural network can adopt advanced machine learning models such as convolutional neural networks (CNNs), recurrent neural networks (RNNs) or their variants (such as long short-term memory networks (LSTMs)). These models can automatically learn the temporal dependencies and spatial structural features in data streams, thereby achieving accurate characterization and classification of fused data stream sets. Taking CNN as an example, the control system can regard the fused data stream set as a multidimensional array (or tensor), which contains voltage values, current values ​​and other related parameters in the time series. CNN extracts local features of data layer by layer through structures such as convolutional layers and pooling layers, and gradually forms a higher level of abstract representation. These representations will eventually be mapped to a preset data stream representation vector set, thereby achieving effective encoding of the fused data stream set. Finally, through the processing of the target data stream recognition neural network, the control system can obtain a representation vector that integrates the characteristics of the target monitoring data stream and its related data streams. This vector not only reflects the overall picture of the current monitoring state, but also contains the accumulation of historical control experience, providing a solid foundation for subsequent control trigger probability prediction, risk assessment, and control strategy optimization.

[0021] As an implementation manner, a preset data flow characterization vector in a preset data flow characterization vector set is a target monitoring data flow characterization vector of the data flow, and the target monitoring data flow characterization vector of the data flow is obtained by downsampling an initial data flow characterization vector of the data flow.

[0022] In the embodiment of the present application, the target monitoring data stream refers to the key data sequences that are continuously monitored and used to evaluate the operating status of the switch cabinet, such as voltage, current, temperature, etc. In order to more effectively represent the information in these data streams, the control system first generates an initial set of data stream characterization vectors. Each vector in this set is an abstract representation of the original data stream data and contains the key feature information of the data stream.

[0023] However, directly using these initial characterization vectors may lead to low computational efficiency due to high dimensionality, and may even introduce unnecessary noise. Therefore, the control system uses downsampling to further optimize these vectors and generate target monitoring data stream characterization vectors. Downsampling is a data dimensionality reduction technology that reduces the number of data points by performing some form of screening or aggregation on the original data, while trying to retain the important features of the original data. In specific implementations, the control system can adopt a variety of downsampling strategies, such as maximum pooling, average pooling, or random sampling. Taking average pooling as an example, assume that there is an initial data stream characterization vector that contains the continuous sampling results of voltage values ​​within a certain period of time. The control system can divide this time period into several subintervals, and then calculate the average value of the voltage value in each subinterval to construct a new, lower-dimensional target monitoring data stream characterization vector. Each element in this new vector represents the average voltage level of the original data stream in the corresponding time period, thereby achieving effective data compression and feature retention. The target monitoring data stream characterization vector generated in this way not only reduces the computational complexity of subsequent processing, but also improves the robustness and generalization ability of the data. In the embodiment of the present application, these vectors will be used as the input of the target data stream identification neural network to achieve accurate identification and analysis of the data associated with the target monitoring data stream in the control data stream set. It should be noted that the specific parameters of the downsampling process (such as the size of the sub-interval, the selection of the pooling function, etc.) often need to be adjusted and optimized according to the actual application scenario and data. The control system can use methods such as cross-validation and grid search to find the best downsampling strategy to ensure that the generated target monitoring data stream representation vector can most accurately reflect the characteristic information of the original data stream. Generating the target monitoring data stream representation vector by downsampling the initial data stream representation vector is an effective data preprocessing and optimization method in the intelligent control system of the intelligent switch cabinet. It helps to improve the efficiency and quality of data processing and provide a solid foundation for subsequent control decisions.

[0024] Step S40: Based on the target data stream identification neural network, search the preset data stream characterization vector set for the preset data stream characterization vector corresponding to each data stream in the fused data stream set, perform characterization vector fusion on each searched preset data stream characterization vector, and obtain the data stream set characterization vector of the fused data stream set; the data stream set characterization vector is used to determine the control trigger probability of the target switch cabinet on the target monitoring data stream.

[0025] In the embodiment of the present application, step S40 involves the application of the target data stream identification neural network and the characterization vector fusion process of the fused data stream set. When the control system performs this step, it makes full use of the data collected and processed in the early stage, as well as the trained target data stream identification neural network, to generate a data stream set characterization vector that can fully reflect the characteristics of the fused data stream set, and then evaluates the control trigger probability of the target switch cabinet on the target monitoring data stream. Specifically, the control system uses the target data stream identification neural network to search in the preset data stream characterization vector set. This preset set contains a large number of preset data stream characterization vectors that can characterize the characteristics of different data streams obtained through historical data analysis. For each data stream in the fused data stream set (including the target monitoring data stream and its related data streams), the control system will try to find the preset data stream characterization vector that best matches it in the preset set. This process may involve algorithms such as vector similarity calculation and nearest neighbor search, aiming to quickly and accurately locate the feature representation corresponding to each data stream. Once the preset data stream characterization vectors corresponding to all data streams are found, the control system will then fuse the characterization vectors. The purpose of fusion is to integrate multiple individual data stream feature vectors into a comprehensive data stream set characterization vector to fully reflect the overall characteristics of the fused data stream set. In actual operation, there may be various fusion methods, including but not limited to weighted averaging, feature concatenation, attention mechanism, etc. For example, the control system can use the attention mechanism to assign different weights according to the importance of each data stream, and then perform weighted summation to obtain the final data stream set representation vector.

[0026] Taking voltage monitoring in power systems as an example, it is assumed that the fused data stream set contains the target monitoring data stream (real-time voltage data) and multiple historical control data streams associated with it (such as data records of circuit breaker actions triggered in similar voltage changes in the past). After the control system finds the preset representation vector corresponding to each data stream through the target data stream recognition neural network, it uses the attention mechanism to weightedly fuse these vectors. In this process, the representation vector of real-time voltage data can be given a higher weight due to its timeliness, while the weight of the historical control data stream may be adjusted according to factors such as its similarity with real-time voltage data and time interval. The final data stream set representation vector not only contains the key features of real-time voltage data, but also integrates the information of historical control experience, providing a comprehensive and in-depth data perspective for the control system. Based on this representation vector, the control system can further analyze the control trigger probability of the target switchgear for the target monitoring data stream (i.e., real-time voltage data). This process may involve another machine learning model (such as a classifier, a regressor, etc.), which predicts the possibility of control triggering based on the data stream set representation vector, thereby providing strong support for the actual control decision of the intelligent switchgear.

[0027] As an implementation method, step S20, determining a related data flow set of a target monitoring data flow in a control data flow set of a target switch cabinet, may specifically include: step S21: extracting a target monitoring data flow characterization vector of the data flow for the target monitoring data flow and the control data flow in the control data flow set of the target switch cabinet; the target monitoring data flow characterization vector is obtained by downsampling an initial data flow characterization vector of the data flow.

[0028] In the embodiment of the present application, the feature extraction of the target monitoring data stream and the control data stream is involved. Specifically, the control system focuses on the target monitoring data stream and each control data stream in the target switch cabinet control data stream set. These data streams may contain the real-time voltage and current values ​​of the power system, the ambient temperature and humidity parameters, or the historical operation records of the switch cabinet. In order to effectively analyze this data, the control system needs to convert it into a form that is easier to process and understand, that is, a data stream representation vector.

[0029] Before extracting the characterization vector of the target monitoring data stream, the control system generates an initial characterization vector of the data stream. This process may involve a variety of data processing and feature extraction techniques, such as signal processing, statistical analysis, or the application of machine learning algorithms. For example, for time series data (such as voltage and current values), the control system can use Fourier transform to extract frequency domain features, or use sliding window technology combined with statistical indicators (such as mean and standard deviation) to generate feature vectors. For non-time series data (such as switch status and environmental parameters), the initial characterization vector may be directly formed through encoding or normalization.

[0030] Since the initial data stream representation vector may contain a large number of data points, directly using it for subsequent analysis is not only computationally intensive, but can also introduce unnecessary noise. Therefore, the control system performs downsampling processing on it to generate a more refined target monitoring data stream representation vector. Downsampling is a data dimensionality reduction technique that reduces the dimension of the vector by reducing the number of data points while retaining the important features of the original data as much as possible.

[0031] The specific method of downsampling can be selected according to the characteristics of the data stream and the analysis requirements. Common downsampling methods include maximum pooling, average pooling, and random sampling. Taking average pooling as an example, the control system divides the initial representation vector into several sub-intervals of equal length, and calculates the average value of the data points in each sub-interval as the element of the target monitoring data stream representation vector at the position of the sub-interval. In this way, through the downsampling process, the control system obtains a target monitoring data stream representation vector with a lower dimension but still rich information.

[0032] Step S22: determining a related data flow set of the target monitoring data flow in the control data flow set of the target switch cabinet according to the spatial similarity coefficient between the target monitoring data flow characterization vectors of the respective target monitoring data flow and the control data flow.

[0033] In an embodiment of the present application, in step S21, the control system has generated a target monitoring data stream characterization vector for each control data stream in the target monitoring data stream and control data stream set. These vectors, as feature representations of the data stream, contain key information of the data stream. In step S22, the control system needs to calculate the spatial similarity coefficient between the target monitoring data stream characterization vector and the control data stream characterization vector to quantify the degree of similarity between them. The spatial similarity coefficient is usually measured by vector distance, and commonly used distance measurement methods include Euclidean distance, Manhattan distance, cosine similarity, etc. In the scenario of intelligent switch cabinets, considering that the feature vector of the data stream may contain multiple dimensions, and the dimensions of different dimensions may be different, the control system is more inclined to use cosine similarity as the basis for calculating the similarity coefficient. Cosine similarity evaluates the similarity of two vectors by measuring the cosine value of the angle between them in the direction. It is not affected by the length of the vector (i.e., the scale of the data stream) and pays more attention to the angular relationship between the vectors.

[0034] After calculating the spatial similarity coefficients between the target monitoring data stream characterization vector and the control data stream characterization vector, the control system determines the set of control data streams related to the target monitoring data stream based on these similarity coefficients. This process usually involves the setting of a similarity threshold. Only when the similarity between the characterization vector of the control data stream and the characterization vector of the target monitoring data stream exceeds this threshold, the control data stream will be considered relevant and included in the set of related data streams. The setting of the similarity threshold depends on the specific application scenario and data analysis requirements. In actual operation, the control system can find the optimal similarity threshold through cross-validation, grid search and other methods to ensure that the selected set of related data streams is neither too broad nor too narrow, and can accurately reflect the potential behavior pattern of the target monitoring data stream.

[0035] For example, suppose that in a power system, the target monitoring data stream is a real-time voltage time series data, and the control data stream set contains multiple voltage time series records that triggered circuit breaker actions under different working conditions in history. The control system has generated a target monitoring data stream representation vector for each record in the target monitoring data stream and the control data stream set.

[0036] In step S22, the control system uses cosine similarity as the basis for calculating the spatial similarity coefficient, traverses each record in the control data flow set, and calculates the cosine similarity between its characterization vector and the target monitoring data flow characterization vector. Subsequently, the control system sets a similarity threshold (for example, 0.8), and screens out all control data flow records with similarities greater than the threshold to form a related data flow set.

[0037] This set of related data streams contains historical control data stream records that are similar or potentially related to the target monitoring data stream in terms of voltage change trends, fluctuation ranges, etc. By analyzing these records, the control system can further predict the possible future behavior patterns of the target monitoring data stream and formulate corresponding control strategies accordingly to ensure the stable operation of the power system.

[0038] As an implementation method, step S21, for the target monitoring data flow and the control data flow in the control data flow set of the target switch cabinet, extracts the target monitoring data flow characterization vector of the data flow, which can specifically include: step S211: generating a current data item set based on the data items of the current data flow, and loading the current data item set into the target downsampling network; the current data flow is the target monitoring data flow or the control data flow in the control data flow set of the target switch cabinet, and the target downsampling network includes an initial data flow characterization vector set.

[0039] In an embodiment of the present application, step S211 is responsible for converting the data stream currently being analyzed (whether it is a target monitoring data stream or a control data stream in the target switch cabinet control data stream set) into a form suitable for subsequent processing, and loading it into a specially designed target downsampling network. The control system first focuses on the "current data stream", which refers to the data stream that is currently being monitored or analyzed. In the scenario of an intelligent switch cabinet, these data streams may contain real-time operating parameters of the power system (such as voltage, current, power factor, etc.), or status information of the switch cabinet (such as circuit breaker position signal, line grounding alarm signal, etc.). These data streams exist in the form of a series of continuous data items, each of which represents the specific value of the data stream at a certain point in time.

[0040] For the needs of subsequent processing, the control system generates a sequence data set based on these data items, namely the "current data item set". This set arranges all data items in the data stream in a certain time sequence or logical order to form an ordered data sequence. For example, in the power system, the current data item set may contain a sequence of voltage values ​​collected every second in the last minute.

[0041] After generating the current set of data items, the control system will then load it into the target downsampling network. The target downsampling network is a pre-trained neural network model specifically designed to process and analyze data streams. This network model contains an "initial data stream representation vector set", which is a database or data structure that stores a large number of initial representation vectors. These initial representation vectors are obtained by analyzing a large amount of historical data stream data, and they can capture the key feature information of different types of data streams.

[0042] When the control system loads the current data item set into the target downsampling network, the network will use the vectors in the initial data stream representation vector set as a reference to further process and analyze the current data item set. Specifically, the network can find the initial representation vector that is most similar or relevant to the current data item set through some form of matching or mapping mechanism, or use these initial representation vectors to guide subsequent feature extraction and dimensionality reduction operations.

[0043] For example, suppose that in a power system, the control system is analyzing a real-time recorded voltage time series data (i.e., the target monitoring data stream). In this time series, each data item represents the voltage value at a certain moment. In order to extract the characteristic information of this data stream, the control system first generates a set of current data items based on the voltage values ​​in the time series, which contains the sequence of all voltage values ​​in the most recent period of time (such as one minute). Subsequently, the control system loads this set of current data items into the target downsampling network. The target downsampling network is a complex model built on a convolutional neural network (CNN) and a self-attention mechanism (such as Transformer). It contains a large number of initial data stream representation vectors, which are obtained by analyzing historical voltage time series data. When the network receives the current data item set, it uses these initial representation vectors as a reference, extracts the deep feature information of the voltage time series through multi-layer convolution and self-attention calculations, and generates a refined target monitoring data stream representation vector for use in subsequent steps.

[0044] Step S212: Based on the target downsampling network, search the initial data stream representation vector corresponding to the data item of the current data stream in the initial data stream representation vector set, enhance the representation information and downsample the representation information of the searched initial data stream representation vector to obtain the target monitoring data stream representation vector of the current data stream.

[0045] In an embodiment of the present application, step S212 is intended to conduct an in-depth analysis of the initial data stream characterization vector through a target downsampling network, and then generate a target monitoring data stream characterization vector that can accurately reflect the characteristics of the current data stream. In step S211, the control system has loaded the data item set of the current data stream into the target downsampling network. At the beginning of step S212, the control system uses the target downsampling network to search in the initial data stream characterization vector set. This set contains a large number of pre-calculated initial characterization vectors that can characterize the characteristics of different types of data streams. The purpose of the search is to find the initial characterization vector that best matches or is similar to the current data stream data item set.

[0046] The search process may involve a variety of algorithms and techniques, such as vector similarity calculation (such as cosine similarity), hash table lookup, etc. The control system calculates the similarity between the current data stream data item set and each vector in the initial data stream representation vector set based on the specific content of the current data stream data item set, and selects the vector with the highest similarity as the matching result. After finding the matching initial data stream representation vector, the control system performs representation information enhancement processing on it. The purpose of this step is to further enhance the feature semantic information of the vector so that it can more accurately reflect the characteristics of the current data stream. Representation information enhancement is usually achieved through complex neural network models, such as Transformer or convolutional neural network (CNN). Taking Transformer as an example, the control system can input the initial representation vector into the Transformer model, and use the self-attention mechanism (Self-Attention) and encoder-decoder (Encoder-Decoder) structure in the model to perform multi-layer nonlinear transformation and feature extraction on the vector. Through this process, the key feature information in the initial representation vector will be enhanced, while redundant or noise information will be suppressed.

[0047] After the characterization information is enhanced, the control system performs downsampling of the characterization information. The purpose of downsampling is to reduce the dimension of the vector to reduce the computational complexity of subsequent processing and remove redundant information. Downsampling can be achieved in many ways, such as mapping (mapping high-dimensional vectors to low-dimensional space), selection (selecting some elements from high-dimensional vectors to form low-dimensional vectors), etc. In actual operation, the control system can adopt a hybrid downsampling method that combines mapping and selection. For example, it can first map the enhanced high-dimensional vector to an intermediate dimensional space by mapping, and then select the most representative elements from it to form the final low-dimensional target monitoring data stream representation vector. For example, suppose that in a power system, the control system is analyzing a real-time recorded voltage time series data (current data stream). In step S211, the control system has generated a set containing the time series data items and loaded it into the target downsampling network.

[0048] In step S212, the control system first uses the network to search for an initial characterization vector that best matches the voltage time series data item set in the initial data stream characterization vector set. It is assumed that the searched initial characterization vector better reflects the basic characteristics of the voltage time series.

[0049] Subsequently, the control system inputs the initial representation vector into the Transformer-based feature enhancement module. Through the Transformer's self-attention mechanism and multi-layer encoder-decoder structure, the key features in the initial representation vector (such as the voltage fluctuation range, change trend, etc.) are significantly enhanced.

[0050] Finally, the control system downsamples the enhanced high-dimensional vector. It uses a hybrid downsampling method to first map the vector to an intermediate dimensional space (such as through a fully connected layer), and then selects the most representative elements (such as through maximum pooling or average pooling operations) to form the final low-dimensional target monitoring data stream representation vector. This vector will serve as the basis for tasks such as data correlation analysis and control strategy formulation in subsequent steps.

[0051] As an implementation mode, the method also includes a training process of a target downsampling network, including: step S21a: obtaining a first set of data items, and loading the first set of data items into an initial downsampling network; the first set of data items is obtained based on a control data flow set of a first a priori switch cabinet, and the initial downsampling network includes an initial data flow characterization vector set.

[0052] In step S21a, the primary task of the control system is to obtain a data set for training, namely, a first set of data items. These data items are not generated out of thin air, but are derived from the control data stream of the switch cabinet in actual operation. Specifically, the control system accesses a database or file system that stores historical control data, and extracts a control data stream set of the first a priori switch cabinet from it. The first a priori switch cabinet refers to the switch cabinet selected as a reference object in the training phase, and its control data stream set records all control behaviors and corresponding data stream changes of the switch cabinet in the past period of time.

[0053] The data items in the control data stream set may include various types of parameters, such as voltage, current, power factor, temperature, humidity, etc. These parameters fluctuate over time and together reflect the operating status of the switchgear. In order to convert these raw data into a form suitable for neural network processing, the control system preprocesses the control data stream set, including data cleaning (removing noise, outliers, etc.), data normalization (scaling the data to a uniform range), and data segmentation (segmenting the continuous data stream into multiple independent data item sets).

[0054] After preprocessing, the control system extracts the first data item set from the control data stream set. This set contains multiple data items, each of which is an ordered combination of a series of parameter values, corresponding to the operating state of the switchgear in a specific time period. For example, in a power system, a data item may contain a sequence of voltage and current values ​​collected every second within a minute.

[0055] Next, the control system loads the first set of data items into the initial downsampling network. The initial downsampling network is a neural network model that has not been fully trained. It is specially designed to process data stream data and extract key feature information from it. This network contains an initial data stream representation vector set, which is a vector set initialized to random values ​​or generated based on some simple rules. These initial representation vectors will serve as the starting point of the network learning process, and will gradually adjust to a state that can accurately represent the characteristics of the data stream as the training progresses.

[0056] In the process of loading a set of data items into the network, the control system formats the data items according to the network input requirements. For example, if the network expects a fixed-length vector sequence as input, the control system can arrange the parameter values ​​in the data items into a vector according to certain rules (such as time order), and may adjust the length of the vector to meet the network requirements by interpolation, truncation, or padding.

[0057] Step S21b: Based on the initial downsampling network, search the initial data stream representation vector set for the initial data stream representation vectors corresponding to each data item in the first data item set that is located before the first distribution position, perform representation information enhancement and representation information downsampling on each searched initial data stream representation vector, obtain an estimated data stream representation vector corresponding to the first distribution position, and obtain a first estimated confidence corresponding to the data item at the first distribution position in the first data item set based on the estimated data stream representation vector corresponding to the first distribution position; the first distribution position is obtained by obtaining the distribution positions of each data item in the first data item set, and the first estimated confidence represents the estimated correlation coefficient between the data stream corresponding to the first distribution position and the preceding data stream set corresponding to the first distribution position.

[0058] In step S21b, the control system uses the initial downsampling network to process each data item in the first data item set to simulate the prediction process of the network in actual applications. Specifically, for each data item in the first data item set, the control system first determines its distribution position in the set, which can be a time point in the time series, an index number in the data sequence, etc., depending on the organization of the data items. Next, the control system searches the initial data stream characterization vector set for the initial characterization vectors corresponding to each data item located before the current data item (i.e., the previous data stream). These initial characterization vectors are previously generated in some way (such as random initialization, simple feature extraction, etc.), and they represent the basic characteristics of the previous data stream.

[0059] After searching for these initial representation vectors, the control system uses the initial downsampling network to further process them. This processing process includes two key steps: representation information enhancement and representation information downsampling. Representation information enhancement aims to enhance the feature semantic information of the vector through complex neural network structures (such as Transformer, CNN, etc.) so that it can more accurately reflect the deep characteristics of the data stream. Representation information downsampling is to reduce the dimension of the vector and reduce the computational complexity while retaining key feature information as much as possible. After these two steps, the control system obtains the estimated data stream representation vector corresponding to the current data item.

[0060] Based on this estimated data stream characterization vector, the control system further calculates the first estimated confidence of the current data item. The first estimated confidence is a quantitative indicator that characterizes the estimated correlation coefficient between the current data stream and the previous data stream set, that is, the degree of association between them. The higher the estimated correlation coefficient, the stronger the correlation between the current data stream and the previous data stream, and the more reliable the network's prediction result.

[0061] In practical applications, the estimated confidence can be used for a variety of purposes. For example, the control system can decide whether to perform corresponding control operations based on the estimated confidence, or adjust the prediction results to better adapt to the actual situation. In addition, the estimated confidence can also be used as an important feedback signal in the network training process to guide the optimization and adjustment of network parameters.

[0062] Step S21c: Obtain a first error according to first estimated confidences corresponding to the data items at each first distribution position in the first data item set.

[0063] In the training process of the target downsampling network, step S21c is a key step in evaluating the prediction performance of the network. It calculates a quantitative error index, namely the first error, based on the network's estimated confidence for each data item in the first data item set. In step S21b, the control system has processed each data item in the first data item set using the initial downsampling network and calculated the first estimated confidence corresponding to each data item. These estimated confidences reflect the network's prediction results of the degree of association between data streams and are an important basis for evaluating network performance. In order to quantify the accuracy of network prediction, the control system needs to calculate an error index, namely the first error. The calculation of the first error is based on the first estimated confidences corresponding to the data items at each first distribution position in the first data item set, and the difference between these estimated confidences and the relationship between the actual data stream (if the actual relationship is known, or indirectly evaluated in some way).

[0064] In actual operation, the control system can adopt a variety of error calculation methods, such as mean square error (MSE), cross-entropy loss, etc. The common goal of these methods is to minimize the difference between the predicted results and the actual results.

[0065] If the actual correlation coefficient is unknown, the control system may need to use an indirect method to evaluate the accuracy of the estimated confidence. For example, it can indirectly evaluate the reliability of the estimated confidence by comparing whether the prediction results of the network in different time periods and different working conditions are consistent and reasonable. For example, in step S21c, the control system first reviews the first estimated confidence calculated in step S21b. These confidences reflect the network's prediction of the degree of correlation between the voltage data at each time point and the voltage data at the previous time point.

[0066] To evaluate the accuracy of these predictions, the control system decided to use mean square error as the error calculation method. It first assumes (or indirectly knows in some way) the actual correlation coefficient between the voltage data at each time point and the voltage data at the previous time point. Then, the control system calculates the square of the difference between the first estimated confidence and the actual correlation coefficient at each time point, and sums and averages these square differences to obtain the first error. If the actual correlation coefficient is completely unknown, the control system can use an alternative solution to evaluate the accuracy of the estimated confidence. For example, it can compare whether the network's prediction results in different seasons and different load conditions are consistent and conform to the basic laws of power system operation. If the network can give reasonable and consistent prediction results under a variety of conditions, then its estimated confidence can be indirectly considered to be reliable.

[0067] Through the calculation in step S21c, the control system obtains a quantized first error value, which will be used to guide the subsequent network parameter optimization process to ensure that the target downsampling network can learn more accurate and effective data flow characterization capabilities.

[0068] Step S21d: Correct the network configuration variables of the initial down-sampling network according to the first error until the first training stop requirement is met, thereby obtaining a target down-sampling network.

[0069] During the training process of the target downsampling network, in step S21d, the control system corrects the network configuration variables of the initial downsampling network, such as weights, biases, and learning rates, according to the first error calculated previously, in order to obtain a target downsampling network with excellent performance after multiple iterations. The following is a detailed explanation of step S21d, which is explained in conjunction with specific application scenarios.

[0070] In step S21c, the control system has calculated the first error based on the first estimated confidence of each data item in the first data item set. This error value reflects the degree of difference between the current network prediction result and the actual result. In order to reduce this difference and improve the prediction accuracy of the network, the control system needs to use this error signal to guide the adjustment of network parameters.

[0071] Specifically, the control system uses the backpropagation algorithm to correct the configuration variables of the network. After multiple iterations of training, when the error value decreases below the preset threshold, or meets other training stop requirements (such as the training round reaches the preset upper limit, the error on the verification set begins to increase, etc.), the control system believes that the network has learned enough data flow representation capabilities, and the network obtained at this time is the target downsampling network.

[0072] The control system uses the back propagation algorithm to calculate the gradient of the error function with respect to the network parameters. In this process, it back propagates the error signal layer by layer and calculates the gradient of the parameters of each layer. Then, the control system updates the values ​​of these parameters according to the preset learning rate (for example, the initial setting is 0.01). In order to speed up the training and ensure stability, the control system also adopts a learning rate decay strategy to gradually reduce the value of the learning rate during the training process. After multiple iterations of training (for example, 100 rounds of training), the control system finds that the error value has been reduced to below the preset threshold (for example, the mean square error is less than 0.01), and the error on the validation set has not shown an increasing trend. At this point, the control system believes that the network has learned enough data flow representation capabilities, so it stops the training process and saves the current network model as the target downsampling network. This trained target downsampling network will be used in subsequent intelligent control tasks, such as real-time prediction of voltage change trends, formulation of control strategies, etc. By continuously learning and optimizing its own parameter configuration, the target downsampling network will be able to better adapt to the complex operating environment of the power system and provide strong support for the intelligent control of intelligent switchgear.

[0073] As an implementation mode, in step S21b, based on the estimated data stream characterization vector corresponding to the first distribution position, a first estimated confidence corresponding to the data item at the first distribution position in the first data item set is obtained, which may specifically include: step S21b1: mapping the estimated data stream characterization vector corresponding to the first distribution position to a preset dimension to obtain an initial data stream distribution characterization vector corresponding to the first distribution position; the initial data stream distribution characterization vector includes characterization vectors corresponding to each candidate data stream in the candidate data stream set.

[0074] In the embodiment of the present application, step S21b1 is a key step in the target downsampling network training process, which involves further processing of the estimated data stream representation vector in order to better analyze the distribution relationship between the data streams. The following is a detailed explanation of step S21b1, which is explained in conjunction with a specific application scenario.

[0075] In this step, the main task of the control system is to convert the estimated data flow characterization vector corresponding to the first distribution position into a new vector, namely the initial data flow distribution characterization vector, through a preset dimension mapping (i.e., linear transformation). This conversion process not only changes the dimension of the vector, but more importantly, it enables the elements in the vector to reflect the distribution characteristics of each candidate data flow in the candidate data flow set in the new feature space. Specifically, the control system first defines a mapping matrix (or weight matrix), the number of rows of which is equal to the number of target dimensions, and the number of columns is equal to the original dimension of the estimated data flow characterization vector. This mapping matrix is ​​preset, and it contains the linear transformation parameters determined by the control system based on prior knowledge or experimental experience. Through matrix multiplication operations, the control system multiplies the estimated data flow characterization vector with the mapping matrix to obtain a new vector, namely the initial data flow distribution characterization vector.

[0076] During the mapping process, each original feature (i.e., each element in the estimated data stream representation vector) is linearly combined into a new feature according to the corresponding parameters in the mapping matrix. This linear combination may retain the key information in the original features, or introduce new feature combinations, thereby helping to reveal more complex correlations between data streams.

[0077] For example, suppose that in an electric power system, the control system is training a target downsampling network to predict the changing trend of the voltage time series. In step S21b, the control system has calculated the estimated data flow characterization vector for each time point based on the historical data in the first data item set. Now, taking a specific time point (i.e., the first distribution position) as an example, the control system needs to perform preset dimension mapping on the estimated data flow characterization vector at that time point. Assume that the original estimated data flow characterization vector is a one-dimensional vector containing multiple parameter features such as voltage, current, and power factor. In order to more comprehensively analyze the distribution relationship of these parameters in the time series, the control system defines a two-dimensional mapping matrix. The number of rows and columns of this matrix corresponds to the target dimension (i.e., the x-axis and y-axis in the two-dimensional space) and the dimension of the original estimated data flow characterization vector, respectively.

[0078] Through matrix multiplication operations, the control system multiplies the original estimated data flow representation vector with the mapping matrix to obtain a new two-dimensional vector, namely the initial data flow distribution representation vector. Each element in this two-dimensional vector is the result of a linear combination of the original features, and they together constitute a point in the two-dimensional feature space at that time point. The relative position relationship between this point and the corresponding points at other time points in the two-dimensional space reflects the distribution characteristics between the voltage data flows at different time points. Through such preset dimensional mapping, the control system can convert the original estimated data flow representation vector into a more intuitive and easy-to-analyze form, providing strong support for subsequent data flow correlation analysis.

[0079] Step S21b2: Normalize the initial data stream distribution characterization vector to obtain the target data stream distribution characterization vector corresponding to the first distribution position; the target data stream distribution characterization vector includes the first estimated confidence corresponding to each candidate data stream in the candidate data stream set, and the candidate data stream set includes the data streams corresponding to each data item in the first data item set.

[0080] In the embodiment of the present application, step S21b2 further extracts a target data flow distribution representation vector that is more intuitive and easy to analyze by normalizing the initial data flow distribution representation vector.

[0081] In this step, the main task of the control system is to normalize the initial data flow distribution characterization vector obtained in step S21b1. Normalization is a common data preprocessing method that eliminates the dimensional differences between different features by scaling the data to a specific interval (usually [0, 1] or [-1, 1]) so that each feature is comparable in subsequent analysis. For the initial data flow distribution characterization vector, the specific method of normalization is to divide each element of the vector by a certain norm of the vector (such as L1 norm, L2 norm, etc.) to ensure that the processed vector has a unit norm. In this way, the initial data flow distribution characterization vector is converted into the target data flow distribution characterization vector.

[0082] Each element of the target data stream distribution representation vector represents the first estimated confidence of the corresponding candidate data stream in the candidate data stream set. The sum of these confidences is 1, reflecting the probability distribution of different candidate data streams being predicted as real data streams under the current prediction task. Due to the normalization process, these confidence values ​​are more intuitive and easy to compare, which facilitates the subsequent data stream correlation analysis.

[0083] For example, suppose that in a power system, the control system is training a target downsampling network to predict the fluctuation trend of the voltage time series. In step S21b1, the control system has obtained the initial data stream distribution representation vector at a certain time point (i.e., the first distribution position) through a preset dimension mapping. The control system normalizes the initial data stream distribution representation vector in step S21b2. Specifically, the L2 norm of the vector is calculated, and then each element of the vector is divided by this norm, thereby obtaining the target data stream distribution representation vector.

[0084] Assume that the initial data stream distribution representation vector is a two-dimensional vector [x, y], where x and y represent the coordinate values ​​of the voltage and current features in the new feature space after mapping. After normalization, the control system obtains the target data stream distribution representation vector [x', y'], where x' and y' are the results of x and y divided by the L2 norm. At this time, the sum of x' and y' is 1, which together reflect the probability distribution of different candidate data streams (such as time series data under different voltage and current combinations) being predicted as real data streams at the current time point.

[0085] Through such normalization processing, the control system not only eliminates the dimensional differences between different features, but also makes the prediction results more intuitive and easy to understand. In the subsequent data flow correlation analysis, the control system can evaluate the similarity and correlation between different candidate data flows based on the confidence value in the target data flow distribution representation vector, thereby formulating a more accurate and effective control strategy.

[0086] Step S21b3: Determine, in the target data stream distribution representation vector, a first estimated confidence level corresponding to a data item at a first distribution position in the first data item set.

[0087] In the embodiment of the present application, in step S21b3, the task of the control system is to directly read or calculate the first estimated confidence corresponding to the data item at the first distribution position in the first data item set in the target data stream distribution representation vector after normalization. This confidence is a numerical value that quantifies the network's confidence in the prediction of the data item as a real data stream. The target data stream distribution representation vector is a multidimensional vector, each dimension (or element) of which represents the first estimated confidence of a candidate data stream in the candidate data stream set. The sum of these confidences is 1, following the principle of probability distribution. In practical applications, the candidate data stream set usually contains all possible data items (or data stream fragments) in the first data item set, and the first distribution position indicates the specific data item position currently being evaluated. Therefore, it is only necessary to find the corresponding element value in the target data stream distribution representation vector according to the first distribution position to obtain the first estimated confidence of the data item. This process is simple and direct, without complex calculations or additional reasoning steps. For example, suppose that in a power system, the control system is training a target downsampling network to predict the voltage value at a specific time point in a voltage time series. The first data item set contains voltage data items at multiple time points in history, and the task of the target downsampling network is to predict the voltage value and its confidence at a certain time point in the future based on the historical data.

[0088] In step S21b2, the control system has obtained the target data stream distribution characterization vector through normalization. This vector is a multidimensional vector, each element of which represents the confidence that the candidate voltage value (i.e., an element in the candidate data stream set) is predicted to be the true voltage value. These candidate voltage values ​​may cover all voltage ranges that appear in the historical data to ensure the comprehensiveness of the prediction. Now, the control system evaluates the voltage data item at a specific time point (i.e., the first distribution position) in the first data item set. According to the index position of this time point in the target data stream distribution characterization vector, the control system directly reads the corresponding element value, i.e., the first estimated confidence of the voltage data item at this time point. Assuming that this confidence value is 0.85, it indicates that the network has 85% confidence that the voltage value prediction at this time point is accurate. Through such steps, the control system can quantitatively evaluate the prediction accuracy of the target downsampling network for each data item, thereby providing important feedback signals for the subsequent training process. If the estimated confidence of a data item is low, the parameters of the network can be adjusted or more training data can be added to improve the overall prediction performance of the network.

[0089] As an implementation mode, the initial downsampling network includes an encoding component, a mapping component, an adjustment component and an output component, the encoding component is used to search for an initial data stream representation vector, the mapping component is used to represent information downsampling, the adjustment component is used to represent information enhancement, and the output component is used to output a first estimated confidence; based on this, step S21d, according to the first error, corrects the network configuration variables of the initial downsampling network until the first training stop requirement is met, and obtains the target downsampling network, including: according to the first error, corrects the network configuration variables of the mapping component, the adjustment component and the output component in the initial downsampling network until the first training stop requirement is met, and obtains the target downsampling network.

[0090] In step S21d of the embodiment of the present application, the initial downsampling network is designed to include four core components: an encoding component, a mapping component, a regulation component, and an output component. Each component undertakes a specific task and works together to achieve effective processing and feature extraction of the data stream.

[0091] Encoding component: responsible for searching for the initial data stream representation vector. During the training process, the encoding component will search for matching representation vectors in the initial data stream representation vector set based on the input data item set. These initial representation vectors serve as the basis for subsequent processing and contain the basic feature information of the data stream.

[0092] Mapping component: used for downsampling of representation information. The mapping component receives the initial representation vector output by the encoding component and performs dimensionality reduction processing on it through a preset linear transformation or nonlinear mapping function to reduce computational complexity and remove redundant information. The vector after dimensionality reduction retains key features and provides a streamlined data foundation for subsequent feature enhancement.

[0093] Adjustment component: responsible for characterizing information enhancement. The adjustment component performs feature enhancement operations on the vector output by the mapping component through a complex neural network structure (such as Transformer, CNN, etc.) to enhance the feature semantic information of the vector so that it can more accurately reflect the deep characteristics of the data stream.

[0094] Output component: used to output the first estimated confidence. The output component receives the feature vector strengthened by the adjustment component and maps it to the first estimated confidence in the form of probability distribution through a function (such as softmax function). These confidence values ​​reflect the network's confidence in predicting different candidate data streams as real data streams.

[0095] In step S21d, the control system corrects the network configuration variables (such as weights, biases, and other parameters) of the mapping component, the adjustment component, and the output component in the initial downsampling network according to the first error (i.e., the difference between the network prediction result and the actual result). This correction process is usually implemented using a back propagation algorithm, which calculates the gradient of the error function with respect to the network parameters and updates the parameter values ​​in the opposite direction of the gradient to reduce the error and optimize the network performance.

[0096] The correction process will be repeated for multiple iterations until the error value is reduced to below the preset threshold, or other training stop requirements are met (such as the training round reaches the upper limit, the verification set error begins to increase, etc.). When these conditions are met, the control system believes that the network has learned enough data flow representation capabilities and saves the network model at this time as the target downsampling network.

[0097] In the embodiment of the present application, the encoding component, mapping component, adjustment component and output component are key components of the target downsampling network. The implementation architecture of the encoding component may be: Hash table lookup: If the initial data stream representation vector set is not very large, a hash table can be used to store these vectors and their corresponding indexes or keys. The encoding component receives the input data item, generates a query key through feature extraction or data conversion, and searches the hash table for a matching initial representation vector.

[0098] Nearest neighbor search: For larger data sets, the encoding component can use an approximate nearest neighbor search algorithm (such as locality sensitive hashing, KD tree, ball tree, etc.) to quickly find the initial representation vector that is most similar to the input data item.

[0099] Neural Network Embeddings: In some cases, the encoding component itself can be a small neural network that takes in an input data item and outputs an embedding vector that is then used to find the closest vector in the initial set of data stream representation vectors.

[0100] The implementation architecture of the mapping component may be: Linear transformation: The mapping component can implement a linear transformation through a weight matrix, which maps the input feature vector to a new low-dimensional space. This usually involves a matrix multiplication operation.

[0101] Nonlinear mapping: In order to handle more complex nonlinear relationships, the mapping component can use kernel methods (such as RBF kernel, polynomial kernel) or nonlinear activation functions in deep neural networks (such as ReLU, sigmoid) to achieve nonlinear mapping.

[0102] Pooling layer: In convolutional neural networks (CNNs), pooling layers (such as max pooling and average pooling) are often used as a means of dimensionality reduction. They can also be seen as a special implementation of the mapping component.

[0103] The architecture of the adjustment component implementation may be: Transformer model: The Transformer model is famous for its self-attention mechanism, which can capture long-distance dependencies in sequence data. The adjustment component can use multiple Transformer encoder layers to enhance the semantic information of the input feature vector.

[0104] Convolutional Neural Network (CNN): For data with local correlation (such as continuous samples in time series data), CNN can extract local features through convolution kernels and gradually enhance feature representation through multi-layer convolution and pooling operations.

[0105] Recurrent Neural Networks (RNNs) and their variants: RNNs (such as LSTM and GRU) are good at processing sequence data and can capture temporal dependencies in the data. The regulation component can use the RNN network to enhance the sequence features.

[0106] The output component implementation architecture may be: Fully connected layer + Softmax function: The output component usually contains one or more fully connected layers to map the feature vector output by the conditioning component to the score space of the target category. The Softmax function is then applied to these scores to produce a first estimated confidence in the form of a probability distribution.

[0107] Logistic regression: In some simple cases, the output component may only contain a logistic regression layer, which directly performs a linear transformation on the feature vector and obtains the predicted probability through the sigmoid function.

[0108] Custom output function: According to the requirements of specific tasks, the output component can also adopt other customized output functions, such as one-to-many strategy in multi-class classification, linear regression model in regression tasks, etc.

[0109] As an implementation mode, the method further includes a process of constructing an initial data stream characterization vector set, including: Step S211a: obtaining a sequence of data items corresponding to each candidate data stream in the candidate data stream set.

[0110] In step S211a of the embodiment of the present application, the control system defines a clear range of candidate data stream sets. This set may include all the switchgear operating status data, control instruction data, and related power system parameter data that have been recorded in history. For example, it may cover the real-time monitoring data of parameters such as voltage, current, power factor, temperature, humidity, etc. under different time periods and different working conditions in the past few years, as well as the corresponding circuit breaker opening and closing records, short-circuit fault processing records, and other control instruction data.

[0111] Next, the control system traverses each candidate data stream in the candidate data stream set. For each candidate data stream, the control system uses corresponding data access and parsing technology to extract the data item sequence according to its data type and storage format. The data item sequence is an orderly arrangement of the data stream in time or space, which records the parameter value changes of the data stream in a specific time period or spatial position.

[0112] Taking the voltage monitoring data stream in the power system as an example, suppose that the candidate data stream set contains a data stream that records the voltage value every minute in the past week. In this data stream, each data item is a voltage value, which are arranged in chronological order and together constitute the data item sequence of the data stream. The control system obtains the data item sequence corresponding to this data stream by reading records in data files, accessing table items in databases, or calling real-time data interfaces.

[0113] In the process of acquiring the data item sequence, the control system also needs to pay attention to the integrity, accuracy and consistency of the data. It can perform preprocessing operations on the data, such as removing outliers, filling missing values, smoothing noise, etc., to ensure the effectiveness of subsequent analysis. At the same time, the control system will also record metadata information such as the length of the data item sequence, sampling frequency, data type, etc. of each candidate data stream, so as to perform unified management and processing in subsequent steps. By executing step S211a, the control system successfully extracts the data item sequence corresponding to each candidate data stream from the candidate data stream set.

[0114] Step S211b: Based on a large-scale neural network, extract representation information of the data item sequence to obtain initial data stream representation vectors corresponding to each candidate data stream.

[0115] In step S211b of the embodiment of the present application, the control system selects a large-scale neural network model suitable for processing sequence data (which has been pre-trained and has mature feature extraction capabilities). Taking into account the complexity and diversity of power system data, the control system can select a neural network model that can capture long-term dependencies and dynamic changes in time series data, such as a long short-term memory network (LSTM), a gated recurrent unit (GRU) or a Transformer. Taking LSTM as an example, the control system uses the data item sequence of each candidate data stream obtained in step S211a as the input of the LSTM network. Each data item sequence contains the power system parameter values ​​(such as voltage, current, power factor, etc.) collected continuously within a specific time period. These data items are arranged in chronological order and together reflect the characteristics of the data stream changing over time.

[0116] In practical applications, the control system can customize and optimize the LSTM network according to the requirements of specific tasks and the characteristics of the data. For example, the performance of the network can be improved by adjusting the network structure (such as increasing or decreasing the number of LSTM layers), changing the activation function, introducing the attention mechanism, etc. In addition, in order to accelerate the training process and improve the generalization ability of the model, the control system can also use batch processing, data enhancement, regularization and other technical means to optimize the training process.

[0117] By executing step S211b, the control system successfully uses a large-scale neural network to extract the initial data stream characterization vector from the data item sequence of the candidate data stream. These characterization vectors not only retain the key feature information of the data stream, but also provide strong support for subsequent tasks such as data correlation analysis and predictive control strategy formulation. In the embodiment of the present application, the successful completion of this step is of great significance for achieving efficient and accurate data-driven control.

[0118] Step S211c: Obtain an initial data stream characterization vector set according to the initial data stream characterization vectors corresponding to the candidate data streams.

[0119] In the embodiment of the present application, step S211c integrates the initial data stream characterization vectors of each candidate data stream generated in step S211b to form a unified set, providing a basis for subsequent data analysis, pattern recognition and control strategy formulation.

[0120] In step S211b, the control system has used a large-scale neural network (such as LSTM, GRU or Transformer, etc.) to conduct in-depth analysis and characterization information extraction on the data item sequence of the candidate data stream, and obtained the initial data stream characterization vector corresponding to each candidate data stream. These vectors are points in a high-dimensional space, and each point carries the key feature information of the corresponding data stream. Next, in step S211c, the control system collects these initial data stream characterization vectors and organizes them into a set according to certain rules, namely, the initial data stream characterization vector set. This set can be regarded as a set of points in a high-dimensional feature space, and each point represents a feature representation of a candidate data stream. In the process of constructing the set, the control system needs to ensure that each initial data stream characterization vector can be added accurately while maintaining the integrity and consistency of the set. To achieve this, the control system can use a data structure (such as a list, array or hash table) to store these vectors, and select a suitable index or key to identify each vector as needed.

[0121] Through step S211c, an initial data stream representation vector set containing rich feature information is constructed, which provides a solid data foundation for subsequent tasks such as data association analysis, pattern recognition, anomaly detection and control strategy formulation.

[0122] As an implementation method, step S22, based on the spatial similarity coefficient between the target monitoring data stream characterization vectors of each of the target monitoring data stream and the control data stream, determines a set of related data streams of the target monitoring data stream in the control data stream set of the target switch cabinet, and specifically may include: step S221: based on the spatial similarity coefficient between the target monitoring data stream characterization vectors of each of the target monitoring data stream and the control data stream, according to the order of increasing spatial similarity coefficients, determines multiple related data streams of the target monitoring data stream in the control data stream set of the target switch cabinet.

[0123] In step S221 of the embodiment of the present application, multiple related data flows that are closest to the behavior pattern of the target monitoring data flow are identified by calculating and comparing the spatial similarity coefficients between the target monitoring data flow and the elements in the control data flow set.

[0124] In step S221, the control system first focuses on the target monitoring data flow characterization vector of each control data flow in the target monitoring data flow and control data flow set. These characterization vectors are obtained through complex feature extraction and dimensionality reduction processing in the previous steps. They exist in the form of points in high-dimensional space, and each point represents the key feature information of the corresponding data flow.

[0125] Next, the control system needs to calculate the spatial similarity coefficient between the target monitoring data stream representation vector and each control data stream representation vector in the control data stream set. The spatial similarity coefficient is a quantitative indicator that measures the degree of proximity between two vectors in spatial position. It reflects the similarity between the two data streams in the feature space. Commonly used methods for calculating spatial similarity coefficients include Euclidean distance, Manhattan distance, cosine similarity, etc. In the scenario of intelligent switchgear, since the control data stream may contain features of multiple dimensions (such as voltage, current, power factor, etc.), and the dimensions and importance of different dimensions may be different, the control system is more inclined to use cosine similarity as the basis for calculating the spatial similarity coefficient.

[0126] The control system traverses each control data flow in the control data flow set, calculates the cosine similarity between its characterization vector and the target monitoring data flow characterization vector, and records these similarity values. Then, the control system sorts the control data flows according to the similarity value from high to low (i.e., the order of increasing spatial similarity coefficient). The result of the sorting is an ordered list of control data flows, and the elements at the front of the list have a higher similarity with the target monitoring data flow.

[0127] Finally, the control system selects multiple control data flows with the highest ranking from this ordered list as relevant data flows for the target monitoring data flow. The specific number of selections depends on the requirements of the actual application scenario and the preset threshold. For example, in a power system, the control system can select the top five control data flows with the highest similarity as relevant data flows, so as to conduct in-depth analysis of the potential correlation between these data flows and the target monitoring data flow.

[0128] Through such steps, the control system can efficiently and accurately filter out multiple relevant data streams that are most similar to the target monitoring data stream from a large amount of historical control data streams, providing strong data support for subsequent control strategy formulation and optimization.

[0129] Step S222: Arrange the relevant data streams in order according to the control intensity of the target switch cabinet to obtain a relevant data stream set of the target monitoring data stream.

[0130] In the embodiment of the present application, step S222 is a key step for further sorting the multiple related data streams screened out in step S221. In this step, the control system sorts the control intensity caused by each related data stream on the target switch cabinet to obtain an ordered set of related data streams.

[0131] In step S221, the control system has screened out multiple related data streams that are closest to the target monitoring data flow behavior pattern based on the spatial similarity coefficient. However, although these data streams are similar to the target monitoring data stream, their actual control impact on the target switch cabinet may be different. Therefore, in step S222, the control system needs to further evaluate the control intensity of these data streams in order to more accurately understand the correlation between them and the target monitoring data stream. Control intensity is a quantitative indicator to measure the degree of influence of data streams on the control behavior of switch cabinets. In the scenario of intelligent switch cabinets, control intensity may be determined by multiple factors, such as the amplitude, rate of change, duration of the data stream, and whether emergency protection measures are triggered. These factors comprehensively reflect the degree and urgency of the impact of the data stream on the operating status of the switch cabinet.

[0132] In order to evaluate the control strength of each relevant data stream, the control system usually consults the control records or logs associated with these data streams. These records contain the response behavior of the switchgear after receiving a specific data stream, such as the opening and closing status of the circuit breaker, short-circuit fault handling, etc. By analyzing these records, the control system can quantitatively evaluate the actual control impact of each relevant data stream on the switchgear. After evaluating the control strength of all relevant data streams, the control system sorts these data streams from high to low according to the control strength (or other order determined according to actual needs). The result of the sorting is an ordered set of relevant data streams, and the front elements in the set have a greater control impact on the switchgear, so they are more closely associated with the target monitoring data stream.

[0133] For example, in a power system, if a related data flow causes an emergency power-off operation of a switchgear, the control intensity of this data flow will be considered to be very high. On the contrary, if a related data flow only reflects a small fluctuation in voltage and does not trigger any control action, the control intensity of this data flow will be relatively low. The control system sorts these data flows according to their control intensity and outputs the sorted results as a set of related data flows of the target monitoring data flow. Through such steps, the control system can not only identify control data flows similar to the target monitoring data flow, but also further sort them according to the actual control impact of these data flows on the switchgear. This helps the control system to more comprehensively understand the correlation between the target monitoring data flow and the historical control data flow, and provide more accurate data support for subsequent control strategy formulation and optimization.

[0134] As an embodiment, the method also includes a training process of a target data flow identification neural network, including: Step S1: obtaining a second set of data items, and loading the second set of data items into an initial data flow identification neural network; the second set of data items is obtained based on a second priori control data flow set of a switch cabinet, and the initial data flow identification neural network includes a preset set of data flow characterization vectors.

[0135] In the embodiment of the present application, step S1 is the starting step of the target data stream recognition neural network training process, which involves the preparation of data and the initialization of the neural network.

[0136] In the data preparation phase, the first task of the control system is to obtain the data set for training, that is, the second data item set. This set is not generated out of thin air, but is obtained through a series of preprocessing steps based on the control data flow set of the second a priori switch cabinet. The second a priori switch cabinet refers to the switch cabinet selected as the reference object in the training phase, and its control data flow set records all the control behaviors and corresponding data flow changes of the switch cabinet in the past period of time.

[0137] The data items in the control data stream set may include time series data of multiple parameters such as voltage, current, power factor, temperature, humidity, etc. These data items are arranged continuously in time and together reflect the operating status and control logic of the switch cabinet. In order to convert these raw data into a form suitable for neural network processing, the control system performs a series of preprocessing operations, such as data cleaning (removing noise and outliers), data normalization (scaling the data to a uniform range), and data segmentation (segmenting the continuous data stream into multiple independent data items).

[0138] After preprocessing, the control system obtains a second set of data items. This set contains multiple data items, each of which is an ordered combination of a series of parameter values, corresponding to the operating status of the switchgear in a specific time period. For example, in a power system, a data item may contain a sequence of voltage and current values ​​collected every second within a minute.

[0139] In the neural network loading phase, after the second set of data items is prepared, the control system then loads it into the initial data flow identification neural network. The initial data flow identification neural network is a neural network model that has not been fully trained. It is specially designed to identify the correlation between data flows and predict the change trend of future data flows. The neural network contains a preset set of data flow representation vectors, which is a set of vectors generated in the preparatory work before training. They extract the key feature information of the historical control data flow in some way (such as the target downsampling network mentioned above). These preset representation vectors serve as prior knowledge of the neural network, which helps the network converge to the optimal solution faster during the training process.

[0140] In the process of loading a set of data items into the neural network, the control system formats the data items according to the requirements of the network input. For example, if the network expects a fixed-length vector sequence as input, the control system can arrange the parameter values ​​in the data items into a vector according to certain rules (such as time order), and may adjust the length of the vector to meet the network requirements by interpolation, truncation or padding. Through such data preparation and neural network loading process, the control system lays a solid foundation for subsequent neural network training. After step S1 is completed, the control system will enter the training phase of the neural network, and improve its ability to identify data flow associations by continuously optimizing network parameters.

[0141] Step S2: Based on the initial data stream identification neural network, search the preset data stream characterization vector set for the preset data stream characterization vectors corresponding to each data item in the second data item set that is located before the second distribution position, perform characterization vector fusion on each searched preset data stream characterization vector, obtain the estimated data stream characterization vector corresponding to the second distribution position, and obtain the second estimated confidence corresponding to the data item at the second distribution position in the second data item set based on the estimated data stream characterization vector corresponding to the second distribution position; wherein the second distribution position is determined among the distribution positions of each data item in the second data item set, and the second estimated confidence represents the estimated correlation coefficient between the data stream corresponding to the second distribution position and the preceding data stream set corresponding to the second distribution position.

[0142] In step S2 of the embodiment of the present application, the control system processes the second data item set using the initial data stream recognition neural network. For each second distribution position in the set (i.e., the position of the data item in the time series or logical sequence), the control system first needs to search the preset data stream representation vector set for the preset data stream representation vectors corresponding to the data items before the position.

[0143] The set of preset data stream representation vectors is generated in the pre-training preparation, which contains key feature information extracted from the historical control data stream. These vectors exist in the form of points in a high-dimensional space, and each point represents a feature representation of the corresponding data stream. The control system searches the set for the preset representation vector that is most similar to the data item before the current second distribution position through a matching mechanism (such as nearest neighbor search based on vector similarity).

[0144] After searching for multiple preset data stream representation vectors, the control system needs to fuse them. Representation vector fusion is a process of merging multiple vectors into a comprehensive vector, which aims to retain the key feature information of each vector and generate a new vector to represent the overall characteristics of the previous data stream set.

[0145] The fusion process may involve a variety of operations, such as vector concatenation, weighted summation, and attention mechanism. Taking weighted summation as an example, the control system can weight each preset representation vector according to its importance (such as the weight calculated based on similarity), and then add the weighted vectors to obtain the estimated data stream representation vector. This vector comprehensively reflects the feature information of the previous data stream set, providing a basis for the subsequent estimated confidence calculation.

[0146] Based on the estimated data stream characterization vector corresponding to the second distribution position, the control system further calculates the second estimated confidence corresponding to the data item at the position. The estimated confidence is a quantitative indicator that characterizes the estimated correlation coefficient between the current data stream and the previous data stream set, that is, the degree of association between them.

[0147] The control system can use a function or model (such as a neural network-based classifier or regressor) to calculate the estimated confidence. This function or model receives the estimated data stream representation vector as input and outputs a value between 0 and 1 as the estimated confidence value. The higher the estimated confidence, the stronger the correlation between the current data stream and the previous data stream, and the more reliable the network's prediction of the data stream.

[0148] For example, suppose that in an electric power system, the control system is training a target data stream recognition neural network to predict the changing trend of the voltage time series. The second data item set contains voltage time series data items in multiple time periods in history, and these data items are arranged in chronological order. For a specific time point in the set (i.e., the second distribution position), the control system first searches for a similar preset voltage data stream representation vector at the previous time point. Then, the control system performs weighted summation of these vectors to obtain an estimated voltage data stream representation vector. Finally, based on this estimated vector, the control system calculates the estimated confidence value corresponding to the voltage data item at the current time point. This value reflects the network's prediction result of the degree of correlation between the current voltage data item and the previous voltage data stream, and provides a reference for the subsequent control strategy formulation.

[0149] Step S3: Obtain a second error according to the second estimated confidences respectively corresponding to the data items at the second distribution positions in the second data item set.

[0150] In the embodiment of the present application, step S3 calculates the error (i.e., loss) by comparing the prediction result of the neural network with the actual situation. This step is crucial for guiding the parameter optimization of the neural network because it provides a direct indicator for measuring the performance of the network. The following is a detailed explanation of step S3, which is explained in conjunction with specific application scenarios.

[0151] When training a target data stream recognition neural network, the goal is to enable the network to accurately predict the association between data streams. In order to evaluate the prediction accuracy of the network, the control system needs to calculate an error value, which reflects the degree of difference between the network's prediction results and the actual results. The smaller the error value, the closer the network's prediction results are to the actual situation, and the better the network's performance.

[0152] In step S2, the control system has calculated the second estimated confidence corresponding to the data item at each second distribution position in the second data item set. These estimated confidences are a quantitative representation of the degree of association between the current data stream and the previous data stream set by the neural network. They reflect the network's confidence in predicting the trend of data stream changes and are an important basis for evaluating the accuracy of network predictions.

[0153] To obtain the second error, the control system first determines an error calculation function or loss function. The loss function is a basic concept in the field of machine learning, which is used to quantify the prediction error of the model. During the training process of the target data stream recognition neural network, the control system can choose one of a variety of loss functions, such as mean square error (MSE), cross-entropy loss, etc.

[0154] The second error value will be used as the basis for subsequent parameter optimization. The control system uses optimization methods such as the back propagation algorithm to adjust the weights and biases of the neural network according to the error value to reduce the error and improve the prediction accuracy of the network. Through multiple iterative training, when the second error value is reduced to below the preset threshold, the control system believes that the network has learned enough data flow recognition capabilities and the training process ends.

[0155] Step S4: Correcting the network configuration variables of the initial data stream identification neural network according to the second error until the second training stop requirement is met, thereby obtaining the target data stream identification neural network.

[0156] In the embodiment of the present application, step S4 corrects the network configuration variables of the neural network according to the second error to optimize the prediction performance of the network until the preset training stop requirements are met, thereby obtaining the final target data stream recognition neural network.

[0157] In step S3, the control system has calculated the second error, which reflects the degree of difference between the predicted result and the actual result of the initial data stream recognition neural network in the current training state. In order to reduce this difference and improve the prediction accuracy of the network, the control system needs to correct the network configuration variables of the network according to the second error. The network configuration variables usually include the weights and biases of the neural network, which are key parameters that determine the behavior of the network. The process of correcting these parameters is usually implemented by the backpropagation algorithm. The backpropagation algorithm is an effective parameter optimization method that uses the chain rule to calculate the gradient of the error function with respect to the network parameters (that is, the derivative of the error with respect to the network parameters), and then updates the network parameters according to the gradient information to reduce the error value.

[0158] During the correction process, the optimization algorithm of Gradient Descent or its variants (such as Stochastic Gradient Descent SGD, Mini-batch Gradient Descent Mini-batch GD, etc.) can be used to minimize the error function by iteratively updating the network parameters until the preset stopping condition is met.

[0159] In order to determine when the training process stops, the control system needs to set a series of training stop requirements. These requirements may include the error value reaching a preset threshold, the training round reaching a preset upper limit, the error on the validation set starting to increase, etc. When any of these requirements is met, the control system believes that the network has learned enough data flow recognition capabilities and the training process can stop.

[0160] For example, suppose that in a power system, the control system is training a target data stream identification neural network to predict the changing trend of the voltage time series. In step S4, the control system first corrects the weights and biases of the neural network according to the second error calculated in step S3. The correction process is implemented by the back propagation algorithm, which uses the gradient descent method to iteratively update the network parameters to minimize the error function. During the training process, the control system continuously monitors the change of the error value and checks whether the preset training stop requirements are met. For example, the control system may set an error threshold (such as 0.01), and when the second error decreases below this threshold, the training process stops. In addition, the control system can also set an upper limit of the training rounds (such as 100 rounds) to prevent the training process from continuing indefinitely. After multiple iterative training, when the training stop requirements are met, the neural network obtained at this time is the target data stream identification neural network.

[0161] As an implementation mode, the initial data stream identification neural network includes an encoding component, a fusion component and an output component, the encoding component is used to search for a preset data stream characterization vector, the fusion component is used to fuse the characterization vector, and the output component is used to output a second estimated confidence; based on this, step S4, correcting the network configuration variables of the initial data stream identification neural network according to the second error until the second training stop requirement is met, and obtaining the target data stream identification neural network, specifically can include: correcting the network configuration variables of the fusion component and the output component in the initial data stream identification neural network according to the second error until the second training stop requirement is met, and obtaining the target data stream identification neural network.

[0162] In the embodiment of the present application, the specific implementation of step S4 is optimized and adjusted for the specific structure of the initial data stream recognition neural network. This neural network is designed to include a coding component, a fusion component and an output component, each of which has a specific task. The coding component and the output component can refer to the aforementioned component structure.

[0163] Encoding component: Responsible for searching for preset representation vectors related to the input data item in the preset data stream representation vector set. These preset representation vectors are generated in the pre-training preparation and contain key feature information of the historical control data stream. The encoding component uses a matching mechanism (such as a search based on vector similarity) to find the preset representation vector that is most relevant to the current input data item.

[0164] Fusion component: After receiving multiple preset representation vectors output by the encoding component, the fusion component is responsible for fusing these vectors. The fusion process may involve multiple operations such as vector concatenation, weighted summation, and attention mechanism, aiming to generate an estimated data stream representation vector that can comprehensively reflect the characteristics of the previous data stream set.

[0165] Output component: Based on the estimated data flow representation vector output by the fusion component, the output component calculates and outputs the second estimated confidence. This confidence is an indicator that quantifies the degree of correlation between the current data flow and the previous data flow set, which is of great significance for subsequent control decisions.

[0166] In step S4, the control system corrects the network configuration variables of the initial data stream recognition neural network according to the second error. However, considering the structural characteristics of the neural network, the control system can selectively optimize the parameters of specific components instead of comprehensively adjusting the entire network.

[0167] Specifically, the control system can prioritize the modification of the network configuration variables of the fusion component and the output component. This is because these two components are directly involved in the generation of the estimated data flow representation vector and the calculation of the second estimated confidence, and their performance has a more direct impact on the final prediction result.

[0168] Correction of fusion components: The control system adjusts the weights and biases in the fusion components according to the gradient information of the second error. The adjustment of these parameters is aimed at optimizing the fusion process so that the generated estimated data stream representation vector can more accurately reflect the characteristics of the previous data stream set.

[0169] Correction of the output component: Similarly, the control system will correct the parameters of the output component based on the gradient information of the second error. The performance of the output component directly affects the accuracy of the second estimated confidence, so optimizing it is a key step to improve the prediction performance of the entire neural network.

[0170] In step S4, the control system corrects the parameters of the fusion component and the output component according to the voltage prediction error (i.e., the second error). Through multiple iterations of training, the control system continuously optimizes the performance of these two components, so that the entire neural network can more accurately predict the changing trend of the voltage time series. When the preset training stop requirements are met, the target data stream recognition neural network obtained by the control system will have a strong voltage prediction capability, providing strong support for the operation of the intelligent switch cabinet.

[0171] As an implementation mode, in step S40, the representation vectors of each searched preset data stream representation vector are fused to obtain a data stream set representation vector of the fused data stream set, which may specifically include: step S41: based on the data stream sorting result of the fused data stream set, the representation vectors of each searched preset data stream are merged to obtain a preset data stream representation vector set.

[0172] In an embodiment of the present application, step S41 merges the searched multiple preset data stream characterization vectors according to the data stream sorting result of the fused data stream set to form an ordered preset data stream characterization vector set. Before executing step S41, the control system has obtained a fused data stream set, which contains multiple data stream fragments related to the target monitoring data stream. These data stream fragments may be time series data of power system parameters such as voltage and current arranged in chronological order, or they may be control instruction records arranged in a certain logical order (such as the order in which events occur). The sorting result of the fused data stream set provides a basis for the subsequent merging of preset data stream characterization vectors.

[0173] In the previous steps, the control system has used the target data stream recognition neural network to process each data stream segment in the fusion data stream set and searched for the preset data stream representation vector corresponding to each data stream segment. These preset representation vectors are generated during the training phase. They extract the key feature information of the historical control data stream in some way (such as the target downsampling network) and are stored in the preset data stream representation vector set as the prior knowledge of the neural network.

[0174] In step S41, the control system reviews the data stream sorting result of the fused data stream set. Then, it will take out the searched preset data stream representation vectors one by one according to the sorting result, and merge them in the same order. The merging process may involve operations such as vector splicing and index correspondence to ensure that the merged preset data stream representation vector set can accurately reflect the original order and content of the fused data stream set.

[0175] For example, suppose that in the power system, the fused data stream set contains the voltage time series data within the last hour, and these data are arranged in chronological order, and the voltage value at each time point corresponds to a preset data stream characterization vector. In step S41, the control system first determines the sorting result of the voltage data within this hour (i.e., the time order), and then takes out and merges the preset characterization vectors at each searched time point one by one according to this time order. The merged preset data stream characterization vector set will be arranged in chronological order, and each vector represents the voltage data characteristics at the corresponding time point.

[0176] Through such a merging process, the control system obtains an ordered set of preset data stream representation vectors, which provides basic data support for the subsequent data stream fusion step. In the subsequent steps, the control system will further process and analyze these preset representation vectors to generate a data stream set representation vector that can fully reflect the characteristics of the fused data stream set.

[0177] Step S42: performing characterization vector fusion on a preset data stream characterization vector set to obtain a data stream fusion characterization vector set; the data stream fusion characterization vector set includes data stream fusion characterization vectors that match the number of data streams in the fused data stream set.

[0178] In the embodiment of the present application, step S42 involves fusing the representation vectors of the preset data stream representation vector set to generate a data stream fusion representation vector set that can comprehensively reflect the characteristics of the fused data stream set.

[0179] In step S41, the control system has merged the searched preset data stream characterization vectors according to the data stream sorting result of the fused data stream set to form an ordered preset data stream characterization vector set. However, each vector in this set only represents the characteristics of a single data stream fragment in the fused data stream set, and cannot fully reflect the characteristics of the entire set. Therefore, the control system needs to combine the characteristic information of these single vectors by means of characterization vector fusion to generate one or more fused characterization vectors that can fully reflect the characteristics of the fused data stream set.

[0180] In the process of characterization vector fusion, the control system needs to select a suitable fusion algorithm. This algorithm should be able to generate one or more new vectors based on the vector content in the preset data stream characterization vector set, and the position of these new vectors in the feature space can accurately reflect the comprehensive characteristics of the original vector. Common fusion algorithms include weighted summation, feature splicing, attention mechanism, etc. Regardless of the fusion algorithm used, the goal of the control system is to generate a set of data stream fusion characterization vectors that matches the number of data streams in the fused data stream set. This means that if the fused data stream set contains N data stream fragments, then the control system should eventually generate a set of N fused characterization vectors. These fused characterization vectors correspond one-to-one to the original data stream fragments in terms of quantity, but in terms of feature information, they integrate the contents of multiple preset characterization vectors.

[0181] Step S43: in the data stream fusion characterization vector set, obtain the data stream fusion characterization vector at the data stream distribution position of the target monitoring data stream as the data stream set characterization vector of the fused data stream set.

[0182] Before executing step S43, the control system has already identified the specific location of the target monitoring data stream in the fused data stream set. This location may be a timestamp (in time series data) or an event identifier (in event-driven data streams). In either case, this location uniquely identifies the distribution of the target monitoring data stream in the fused data stream set.

[0183] In step S42, the control system generates a data stream fusion representation vector set by fusing the preset data stream representation vector set. Each vector in this set is a comprehensive representation of the features of multiple data stream segments at a specific position (such as a certain time point or event) in the fused data stream set. Since the fusion process maintains the order and correspondence of the original data stream, the vectors in the data stream fusion representation vector set are also arranged in the original order of the fused data stream set.

[0184] In step S43, the control system searches for the corresponding fusion representation vector in the data stream fusion representation vector set according to the data stream distribution position of the target monitoring data stream. Since the vectors in the set correspond one-to-one to the data stream segments in the fusion data stream set, the control system can directly find the fusion representation vector corresponding to the target monitoring data stream through the position index.

[0185] This fusion representation vector integrates the characteristic information of the target monitoring data stream and its related data streams, and reflects the overall characteristics of the fusion data stream set at the distribution position of the target monitoring data stream. Therefore, the control system uses this vector as the data stream set representation vector of the fusion data stream set for subsequent control decision-making or data analysis process.

[0186] In step S43, the control system needs to find the data stream fusion characterization vector corresponding to the target monitoring data stream (i.e., the voltage time series data recorded in real time). Since the target monitoring data stream corresponds to a specific time period in the fused data stream set, the control system can directly find the corresponding fusion characterization vector in the time period through the time index. This vector integrates the characteristic information of all voltage data streams in the last minute, and reflects the overall change trend and characteristics of the voltage time series data in the time period. Therefore, the control system uses this vector as the data stream set characterization vector of the fused data stream set for subsequent voltage prediction, control strategy formulation and other tasks.

[0187] As an embodiment, the method further includes:

[0188] Step S100: obtaining a data flow characterization vector of a target monitoring data flow and a switch cabinet characterization vector of a target switch cabinet;

[0189] Step S200: Load the data flow set characterization vector, the data flow characterization vector of the target monitoring data flow, and the switch cabinet characterization vector of the target switch cabinet into the target estimation neural network to obtain the control trigger probability of the target switch cabinet on the target monitoring data flow.

[0190] In the embodiment of the present application, step S100 involves the acquisition of the data stream characterization vector of the target monitoring data stream. In this part, the primary task of the control system is to obtain the data stream characterization vector of the target monitoring data stream. The target monitoring data stream is usually the power system parameter data collected in real time, such as voltage, current, power factor, etc. These data fluctuate over time and reflect the operating status of the power system. In order to convert these raw data into characterization vectors that can be understood by the neural network, the control system adopts a series of feature extraction techniques. These techniques may include statistical feature calculation (such as mean, standard deviation, maximum value, minimum value, etc.), time-frequency domain analysis (such as Fourier transform, wavelet transform, etc.), or automatic feature extraction through pre-trained deep learning models (such as convolutional neural network CNN, recurrent neural network RNN ​​and its variants LSTM, GRU, etc.). Taking voltage monitoring in the power system as an example, the control system can first calculate the mean and standard deviation of the voltage time series to reflect the overall level and fluctuation of the voltage. Then, it can use Fourier transform to convert the voltage signal from the time domain to the frequency domain, and extract the energy distribution of different frequency components as features. In addition, if there is a pre-trained deep learning model in the system, the control system can also directly input the voltage time series into the model, allowing the model to automatically extract deep feature representations.

[0191] Finally, the control system will obtain a fixed-length numerical vector, which is the data flow representation vector of the target monitoring data flow. It contains the key information in the original data and is the basis for subsequent control decisions.

[0192] Then the switchgear characterization vector of the target switchgear is obtained. Similar to the target monitoring data stream, the control system also needs to obtain the switchgear characterization vector of the target switchgear. This vector reflects various attribute information of the switchgear, such as physical characteristics, operating status, historical control records, etc. In order to construct the switchgear characterization vector, the control system collects relevant data of the switchgear and processes it appropriately. These data may include static parameters such as the model, capacity, rated voltage and current of the switchgear, and may also include dynamic parameters such as the current load rate, temperature, humidity, and fault data. In addition, the control system can also refer to the historical control records to understand the control behavior and response of the switchgear in the past period of time. After obtaining these data, the control system selects the appropriate characterization method according to the specific application scenario and requirements. For static parameters, the control system can directly convert them into elements in the vector; for dynamic parameters and historical control records, the control system can use more complex data processing methods, such as time series analysis, cluster analysis, or feature extraction through additional machine learning models.

[0193] Finally, the control system will obtain another fixed-length numerical vector, namely the switchgear characterization vector of the target switchgear. This vector, together with the data stream characterization vector of the target monitoring data stream, constitutes the input feature set required for subsequent control decisions.

[0194] For example, suppose that in a power system, the control system is monitoring and controlling an intelligent switchgear in real time. The target monitoring data stream is a real-time collected voltage time series data, and the target switchgear is the key equipment responsible for the voltage regulation of the power line.

[0195] In step S100, the control system first converts the voltage time series into a data stream representation vector through feature extraction technology (such as statistical feature calculation and deep learning model automatic feature extraction). This vector contains key information such as the overall level of voltage, fluctuation, and frequency component distribution.

[0196] At the same time, the control system also collects relevant data of the target switchgear, including static parameters such as its model, capacity, rated voltage and current, as well as dynamic information such as fault data, control behavior and response status in the past period of time. By processing and analyzing these data, the control system constructs a switchgear characterization vector of the target switchgear. This vector reflects the physical characteristics, operating status and historical control experience of the switchgear.

[0197] Finally, the control system obtains two characterization vectors: one represents the key features of the target monitoring data stream, and the other represents the comprehensive attributes of the target switchgear. These two vectors will serve as input feature sets for the subsequent control decision process, providing strong support for the intelligent control of the intelligent switchgear.

[0198] Step S200 uses the target prediction neural network to process the input feature vector to predict the control trigger probability of the target switch cabinet for the target monitoring data stream. The target prediction neural network is a supervised debugging prediction model that can predict the control trigger probability based on the input feature vector. In the training phase, the control system uses a set of debugging learning sample tuples containing prior labels to debug the initial prediction neural network. These tuples are composed of the data stream representation vector of the sample data stream, the switch cabinet representation vector of the prior switch cabinet, and the fusion representation vector between them. Each tuple corresponds to a priori label, that is, the control trigger probability of the target switch cabinet for the target monitoring data stream under specific conditions. By minimizing the error between the predicted probability and the priori label, the control system continuously adjusts the parameters of the neural network until the model performance meets the preset standard, thereby obtaining the target prediction neural network. This network has powerful feature extraction and pattern recognition capabilities, and can accurately reflect the control behavior laws of the target switch cabinet under different conditions.

[0199] In the real-time control stage, the control system first obtains the data stream characterization vector of the target monitoring data stream and the switch cabinet characterization vector of the target switch cabinet. These data are obtained through the feature extraction technology mentioned above, and contain the key feature information of the target monitoring data stream and the comprehensive attribute information of the target switch cabinet. Then, the control system loads the data stream set characterization vector generated in step S40 together with the above two characterization vectors into the target estimation neural network. The data stream set characterization vector comprehensively reflects the overall characteristics of the fused data stream set, and it, together with the data stream characterization vector of the target monitoring data stream and the switch cabinet characterization vector of the target switch cabinet, constitutes the input feature set of the neural network.

[0200] After receiving the input feature set, the target estimation neural network will process the feature vector using its internal complex network structure and parameter settings. This processing may include multiple operations such as multi-layer nonlinear transformation, activation function application, feature fusion, etc., aiming to extract deep features in the input data and identify patterns related to the target switchgear control behavior.

[0201] After a series of data processing, the target estimation neural network finally outputs a value between 0 and 1, which is the control trigger probability of the target switchgear on the target monitoring data stream. This probability value reflects the possibility of the target switchgear executing a control action (such as adjusting voltage, switching circuit breaker status, etc.) in the current state.

[0202] For example, suppose that in the power system, the target monitoring data stream is a real-time recorded voltage time series data, and the target switch cabinet is an intelligent device responsible for the voltage regulation of the power line. In step S200, the control system first obtains the data stream characterization vector of the voltage time series and the switch cabinet characterization vector of the switch cabinet. Then, it loads these data together with the data stream set characterization vector generated in step S40 into the target estimation neural network.

[0203] By processing the input feature vector, the neural network identifies the potential correlation between the current voltage fluctuation and the historical control behavior of the target switchgear. Based on this correlation, the neural network predicts the probability of the target switchgear performing voltage regulation control actions at the current voltage level. If the predicted probability is higher than the preset threshold, the control system will consider it necessary to adjust the voltage to maintain the stable operation of the power system. Therefore, it will issue corresponding control instructions based on the prediction results to command the target switchgear to perform the corresponding voltage regulation operation. Conversely, if the predicted probability is lower than the threshold, the control system may choose to keep the current state unchanged or perform other forms of monitoring and evaluation.

[0204] An embodiment of the present application provides a control system, including a CPU main control chip (domesticated) and a FLASH memory chip (domesticated), etc. The FLASH memory chip can run a computer program on a processor, and the CPU main control chip implements some or all of the steps in the above-described method when executing the program.

[0205] Figure 2 A schematic diagram of a hardware entity of a control system provided in an embodiment of the present application, such as Figure 2 As shown, the hardware entity of the control system 1000 includes: a CPU main control chip 1001 and a FLASH storage chip 1002, wherein the FLASH storage chip 1002 stores a computer program that can be run on the CPU main control chip 1001, and when the CPU main control chip 1001 executes the program, the steps in the method of any of the above embodiments are implemented.

[0206] The above is only an implementation method of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application.

Claims

1. An intelligent control method for an intelligent switch cabinet, It is characterized in that The method comprises: Obtain a target monitoring data stream and a target switch cabinet control data stream set; Determining a related data flow set of the target monitoring data flow in the control data flow set of the target switch cabinet; According to the relevant data stream set and the target monitoring data stream, a fused data stream set is obtained, and the fused data stream set is loaded into a target data stream identification neural network; the target data stream identification neural network includes a preset data stream characterization vector set, and the target data stream identification neural network is obtained by debugging the control data stream set of the prior switch cabinet; Based on the target data stream identification neural network, the preset data stream characterization vector set is searched for the preset data stream characterization vector corresponding to each data stream in the fused data stream set, and the characterization vectors of each searched preset data stream characterization vector are fused to obtain the data stream set characterization vector of the fused data stream set; the data stream set characterization vector is used to determine the control trigger probability of the target switch cabinet on the target monitoring data stream; Wherein, determining the relevant data flow set of the target monitoring data flow in the control data flow set of the target switch cabinet includes: Extracting a target monitoring data stream characterization vector of the data stream for the target monitoring data stream and the control data stream set of the target switch cabinet; the target monitoring data stream characterization vector is obtained by downsampling an initial data stream characterization vector of the data stream; Determine a relevant data stream set of the target monitoring data stream in the control data stream set of the target switch cabinet according to a spatial similarity coefficient between respective target monitoring data stream characterization vectors of the target monitoring data stream and the control data stream; The step of extracting a target monitoring data flow characterization vector of the data flow from the target monitoring data flow and the control data flow set of the target switch cabinet includes: Generate a current data item set according to the data items of the current data flow, and load the current data item set into a target downsampling network; the current data flow is the target monitoring data flow or the control data flow in the control data flow set of the target switch cabinet, and the target downsampling network includes an initial data flow characterization vector set; Based on the target downsampling network, searching the initial data stream representation vector set for the initial data stream representation vector corresponding to the data item of the current data stream, performing representation information enhancement and representation information downsampling on the searched initial data stream representation vector, and obtaining the target monitoring data stream representation vector of the current data stream; The step of fusing the representation vectors of the searched preset data stream representation vectors to obtain the data stream set representation vector of the fused data stream set includes: According to the data stream sorting result of the fused data stream set, merging the searched preset data stream representation vectors to obtain a preset data stream representation vector set; Performing characterization vector fusion on the preset data stream characterization vector set to obtain a data stream fusion characterization vector set; the data stream fusion characterization vector set includes data stream fusion characterization vectors matching the number of data streams of the fused data stream set; From the data stream fusion characterization vector set, obtain a data stream fusion characterization vector at a data stream distribution position of the target monitoring data stream as a data stream set characterization vector of the fused data stream set; The preset data flow characterization vector in the preset data flow characterization vector set is a target monitoring data flow characterization vector of the data flow, and the target monitoring data flow characterization vector of the data flow is obtained by downsampling the initial data flow characterization vector of the data flow; The method further comprises: Acquire a data stream characterization vector of the target monitoring data stream and a switch cabinet characterization vector of the target switch cabinet; The data stream set characterization vector, the data stream characterization vector of the target monitoring data stream, and the switch cabinet characterization vector of the target switch cabinet are loaded into the target estimation neural network to obtain the control trigger probability of the target switch cabinet on the target monitoring data stream.

2. The method according to claim 1, It is characterized in that The method further comprises: Acquire a first set of data items, and load the first set of data items into an initial downsampling network; the first set of data items is obtained based on a control data flow set of a first a priori switch cabinet, and the initial downsampling network includes the initial data flow characterization vector set; Based on the initial downsampling network, the initial data stream representation vector set is searched for the initial data stream representation vector corresponding to each data item located before the first distribution position in the first data item set, and the representation information of each searched initial data stream representation vector is enhanced and down-sampled to obtain an estimated data stream representation vector corresponding to the first distribution position, and according to the estimated data stream representation vector corresponding to the first distribution position, a first estimated confidence corresponding to the data item at the first distribution position in the first data item set is obtained; the first distribution position is obtained by obtaining the distribution positions of each data item in the first data item set, and the first estimated confidence represents the estimated correlation coefficient between the data stream corresponding to the first distribution position and the preceding data stream set corresponding to the first distribution position; Obtaining a first error according to first estimated confidences respectively corresponding to the data items at the first distribution positions in the first data item set; The network configuration variables of the initial down-sampling network are corrected according to the first error until a first training stop requirement is met, thereby obtaining the target down-sampling network.

3. The method according to claim 2, It is characterized in that The obtaining, according to the estimated data stream characterization vector corresponding to the first distribution position, a first estimated confidence corresponding to the data item at the first distribution position in the first data item set comprises: Performing preset dimension mapping on the estimated data stream characterization vector corresponding to the first distribution position to obtain an initial data stream distribution characterization vector corresponding to the first distribution position; the initial data stream distribution characterization vector includes characterization vectors corresponding to each candidate data stream in the candidate data stream set; Normalizing the initial data stream distribution characterization vector to obtain a target data stream distribution characterization vector corresponding to the first distribution position; the target data stream distribution characterization vector includes first estimated confidences corresponding to each candidate data stream in the candidate data stream set, and the candidate data stream set includes data streams corresponding to each data item in the first data item set; In the target data stream distribution representation vector, determining a first estimated confidence corresponding to a data item in the first data item set at the first distribution position; The initial downsampling network includes an encoding component, a mapping component, an adjustment component and an output component, wherein the encoding component is used to search for an initial data stream representation vector, the mapping component is used to represent information downsampling, the adjustment component is used to represent information enhancement, and the output component is used to output a first estimated confidence; The method of correcting the network configuration variables of the initial down-sampling network according to the first error until a first training stop requirement is met to obtain the target down-sampling network comprises: The network configuration variables of the mapping component, the regulating component and the output component in the initial downsampling network are corrected according to the first error until a first training stop requirement is met, thereby obtaining the target downsampling network.

4. The method according to claim 3, It is characterized in that The method further comprises: Obtaining a data item sequence corresponding to each candidate data stream in the candidate data stream set; Based on a large-scale neural network, extracting representation information of the data item sequence to obtain initial data stream representation vectors corresponding to each candidate data stream; Obtaining the initial data stream characterization vector set according to the initial data stream characterization vectors respectively corresponding to the candidate data streams; The step of determining a set of related data streams of the target monitoring data stream in the set of control data streams of the target switch cabinet according to a spatial similarity coefficient between respective target monitoring data stream characterization vectors of the target monitoring data stream and the control data stream comprises: According to the spatial similarity coefficient between the target monitoring data stream characterization vectors of the target monitoring data stream and the control data stream, and according to the order of increasing spatial similarity coefficients, a plurality of related data streams of the target monitoring data stream are determined in the control data stream set of the target switch cabinet; Arrange the relevant data streams in sequence according to the control intensity of the target switch cabinet to obtain a relevant data stream set of the target monitoring data stream.

5. The method according to claim 1, It is characterized in that The method further comprises: Acquire a second set of data items, and load the second set of data items into an initial data flow identification neural network; the second set of data items is obtained based on a second a priori control data flow set of the switch cabinet, and the initial data flow identification neural network includes the preset data flow characterization vector set; Based on the initial data stream identification neural network, searching the preset data stream characterization vector set for a preset data stream characterization vector corresponding to each data item located before the second distribution position in the second data item set, performing characterization vector fusion on each searched preset data stream characterization vector to obtain an estimated data stream characterization vector corresponding to the second distribution position, and obtaining a second estimated confidence corresponding to the data item at the second distribution position in the second data item set according to the estimated data stream characterization vector corresponding to the second distribution position; The second distribution position is determined in each data item distribution position of the second data item set, and the second estimated confidence represents an estimated correlation coefficient between a data stream corresponding to the second distribution position and a preceding data stream set corresponding to the second distribution position; Obtaining a second error according to second estimated confidences respectively corresponding to the data items at each second distribution position in the second data item set; The network configuration variables of the initial data flow identification neural network are corrected according to the second error until a second training stop requirement is met, thereby obtaining the target data flow identification neural network.

6. The method according to claim 5, It is characterized in that The initial data stream identification neural network includes an encoding component, a fusion component and an output component, wherein the encoding component is used to search for a preset data stream representation vector, the fusion component is used for representation vector fusion, and the output component is used to output a second estimated confidence; The method of correcting the network configuration variables of the initial data flow identification neural network according to the second error until a second training stop requirement is met to obtain the target data flow identification neural network comprises: The network configuration variables of the fusion component and the output component in the initial data flow identification neural network are corrected according to the second error until the second training stop requirement is met, thereby obtaining the target data flow identification neural network.

7. A control system comprising a memory and a processor, wherein the memory stores a computer program executable on the processor, It is characterized in that When the processor executes the program, the steps in the method according to any one of claims 1 to 6 are implemented.

Citation Information

Patent Citations

  • Power monitoring system based on big data fusion model

    CN113050523A

  • Intelligent high-voltage switch cabinet fault diagnosis method and system based on heterogeneous graph structure learning

    CN116244617A